# FLAME AI Decision Audit System — Ethics Framework & Rubric Equity Statement
**Rodrik Publishing | Version 1 | Effective: May 1, 2026**

## Purpose

This document states the ethical principles that govern the FLAME AI Decision Audit System, acknowledges known limitations in the scoring rubric, and documents the steps being taken to address them. It is published publicly as part of FLAME's commitment to transparency over false confidence.

---

## Five Core Ethical Principles

**1. Do No Harm by Design**
The harm detection classifier runs on every user input at intake and per dimension. Decisions describing discriminatory intent against protected groups trigger FLAME Blocked before any scoring occurs. This is the only non-negotiable principle in the system — it cannot be overridden by the user.

**2. Transparency Over False Confidence**
The scoring rubric penalizes vague and undocumented claims regardless of intent. The public self-audit log publishes honest scores including regressions. The report disclaimer states explicitly that FLAME is not a regulatory standard and results are AI-generated. The tool is designed to never project more authority than it has earned.

**3. Human Judgment as the Final Check**
Every report requires qualified human review before any consequential decision is made. The appeals process routes disputes to human judgment. The tool informs decisions — it does not make them.

**4. Accountability by Design**
The Leadership dimension requires a named accountable individual with documented authority. A governance structure with no named accountable person scores low regardless of how sophisticated the other dimensions are. The Triage Protocol v3, credential escrow, and published governance documents exist to ensure someone is visibly responsible for what the tool does.

**5. Governance Reality Over Governance Theater**
The rubric rewards specific named mechanisms, documented processes, and tested safeguards. Stated intentions score 5 or below. Documented and tested mechanisms score 15 or above. The rubric is calibrated to distinguish organizations that think about governance from those that practice it.

---

## Rubric Equity Statement

FLAME acknowledges the following known limitations in its scoring rubric:

### Western Governance Bias
The FLAME rubric was designed by one person with one cultural background. The scoring criteria reward:
- Individual accountability structures over collective accountability models
- Formal written documentation over oral governance traditions
- Quantitative evidence standards over relationship-based oversight approaches
- English-language articulation of governance concepts

Leaders from non-Western governance contexts, collective leadership traditions, or oral accountability cultures may score lower on the current rubric not because their governance is weaker but because it is less legible to the framework's design assumptions. This is a known limitation, not a feature.

### Claude Model Bias
The scoring engine runs on Claude by Anthropic. Any systematic tendencies in Claude's underlying model — favoring certain writing styles, cultural frameworks, or documentation conventions — are embedded in every FLAME score. No audit of Claude's behavior within the FLAME context has been conducted.

---

## Steps Being Taken to Address Rubric Equity

**Completed:**
- Spanish-language interface deployed with auto-detection — expanding accessibility to Spanish-speaking leaders
- Bilingual support for all dimension questions, transitions, consent flows, and reports

**In Progress:**
- Methodology review outreach sent to AI Now Institute and Algorithmic Justice League — seeking external critique of the rubric from researchers working on algorithmic accountability
- Three targeted external conversations planned before Q3 2026:
  1. One AI ethics researcher from a non-Western governance context
  2. One public sector practitioner from a developing country
  3. One labor rights perspective reviewing the Morality and Ethics dimension questions

**Planned:**
- Demographic logging in session data — capturing country and organizational size to detect whether certain user types score systematically lower
- Consistency testing — publishing variance data to establish scoring stability as a prerequisite for equity analysis
- External methodology review — written summary of findings to be published publicly

**Timeline:** All planned steps to be initiated before Q3 2026. Findings from external conversations to be incorporated into rubric v2.

---

## Appeals and Rubric Challenges

Users who believe their score was inaccurate may appeal per the Triage Protocol v3 appeals process. Christian Calix serves as secondary reviewer with a binding recommendation within 14 days.

Users who believe the rubric itself systematically disadvantaged their governance approach — rather than misapplied it — may raise this concern at flame@rodrik.io. These concerns are logged and inform the rubric equity improvement process. They are not resolved through the individual appeals process but through the external methodology review.

---

## Review Cadence

This document is reviewed and updated:
- Following each completed external conversation or methodology review
- When rubric changes are made based on equity findings
- Annually as part of the FLAME governance review cycle

*Version 1 — Initial publication. Acknowledges Western governance bias and Claude model bias. Documents completed, in-progress, and planned equity improvements.*
*Published: May 1, 2026*